Evaluating and Characterizing Human Rationales

By Samuel Carton, Anirudh Rathore, Chenhao Tan. 2020

56 citations recorded by the source corpus.

Two main approaches for evaluating the quality of machine-generated rationales are: 1) using human rationales as a gold standard; and 2) automated metrics based on how rationales affect model behavior. An open question, however, is how human rationales fare with these automatic metrics. Analyzing a variety of datasets and models, we find that human rationales do not necessarily perform well on these metrics. To unpack this finding, we propose improved metrics to account for model-dependent baseline performance. We then propose two methods to further characterize rationale quality, one based on model retraining and one on using "fidelity curves" to reveal properties such as irrelevance and redundancy. Our work leads to actionable suggestions for evaluating and characterizing rationales.

Subfields: Artificial Intelligence, Computer Science Applications, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Science Applications

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

Few-Shot Self-Rationalization with Natural Language Prompts (2021, neutral): “This result is in line with Carton et al. (2020) who also recommend breaking down the evaluation of explanations w.r.t. labels whenever possible.”
On the Diversity and Limits of Human Explanations (2021, background): “We emphasize that these different forms of questions entail different properties of the collected human explanations, as evidenced by Carton et al. (2020).”
Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processing (2021, background): “[23] expect that FEVER highlights are non-comprehensive, in contrast to DeYoung et al.”
Teach Me to Explain: A Review of Datasets for Explainable NLP (2021, background): “(implies the prediction, §3; first example in Table 2) and comprehensive (its complement in the input does not imply the prediction, §3; second example in Table 2) is regarded as faithful to the prediction it explains (DeYoung et al., 2020a; Carton et al., 2020).”